revise deploy.prototxt

[xizi]

Hi, shicai, can i directly revise deploy.prototxt to get trainval.prototxt and what is the need to pay attention to detail?

[shicai]

yes.

• add data layers, and loss/accuracy layers
• add param { lr_mult: 1 decay_mult: 1} for conv layers
• add param { lr_mult: 0 decay_mult: 0} for 3 times for each batchnorm layer
• add fillers for conv and scale layers if not finetuing

[yuanyuanli85]

I think you can refer to code to generate deploy and train_val prototxt. https://github.com/liuzhuang13/DenseNetCaffe/ or https://github.com/yuanyuanli85/DenseNet-Caffe

[dlyldxwl]

请问一下，为什么bn层需要加3次呢，不是只有两个参数吗，而且这两个参数也是不用学习的呀

[shicai]

caffe默认的batchnorm层有3个参数，正是因为这些参数是需要计算而不是需要训练的，所以才需要将lr_mult和decay_mult设置为0。

[dlyldxwl]

谢谢回答！我看了一下源码，第三个参数大概是滑动系数和的意思，前两个分别是mean的和、var的和。但是我看了一下你的densenet model为啥第三个系数值是1呢，你做了reshape吗？

[dlyldxwl]

错了，我可视化的是你的nobilenet model的conv1/bn里的第三个参数值是1。你训练的model很赞，给力！

[shicai]

thx~